基于深度学习的道路小目标检测改进算法  

Improved Algorithm for Small Road Object Detection Based on Deep Learning

作　　者：关浩 关宇 王宇峰 张宏[3] GUAN Hao;GUAN Yu;WANG Yufeng;ZHANG Hong(Inner Mongolia Autonomous Region Transportation Development Center,Hohhot 010020,China;Hohhot City Transportation Service Center,Hohhot 010010,China;Institute of Transportation,Inner Mongolia University,Hohhot 010070,China)

机构地区：[1]内蒙古自治区交通运输事业发展中心,呼和浩特010020 [2]呼和浩特市交通运输服务中心,呼和浩特010010 [3]内蒙古大学交通学院,呼和浩特010070

出　　处：《内蒙古大学学报(自然科学版)》2025年第1期77-87,共11页Journal of Inner Mongolia University:Natural Science Edition

基　　金：国家自然科学基金项目(62362053);内蒙古高等学校青年科技英才支持计划项目(NJYT23060);“内蒙古科技成果转移转化示范区、高校协同创新基地和高校创业实训基地”建设项目(超级算力项目)(21300-231510);2024年度基础研究与应用基础研究(2024-规-基-29)。

摘　　要：针对现有道路目标检测技术在小目标识别方面的挑战,如精度低、误检率高和漏检问题,提出一种基于YOLOv5的目标检测改进方法(GCE-YOLOv5),旨在提高小目标检测能力。首先,在主干网络上集成全局注意力机制;其次,在头部网络中增加小目标检测层;最后,将原来的损失函数改为高效交并比损失函数。在KITTI数据集上评估了GCE-YOLOv5算法,并将性能与YOLOv5算法进行了比较。结果表明,GCE-YOLOv5将检测mAP@0.5提高了1.6%,P提高了1.5%。与YOLOv5算法相比,GCE-YOLOv5模型检测道路小目标的精度更高,有效改善了小目标漏检问题,具有更好的检测效果。模型检测速度满足实时性的要求,适合集成到自动驾驶系统中,用于提升车辆在复杂道路环境中的感知能力。To cope with the challenges of existing techniques in small road object recognition,such as low accuracy,high false detection rate and missed detection,a YOLOv5-based object detection improvement method(GCE-YOLOv5)is proposed,aiming to improve the small object detection capability.Firstly,the Global Attention Mechanism(GAM)module is integrated on the backbone network;secondly,a small object detection layer is added to the header network;and finally,the original loss function is changed to Efficient Intersection over Union(EIoU)loss.We evaluated the performance of the GCE-YOLOv5 algorithm on the KITTI datasets,and compared it with the YOLOv5 algorithm.The results show that the GCE-YOLOv5 algorithm improves mAP@0.5 by 1.6%and P by 1.5%.Compared with the YOLOv5 algorithm,the GCE-YOLOv5 model has higher accuracy in detecting small road object,which effectively improves the problem of missed detection of small object and has better detection performance.The GCE-YOLOv5 model detection speed meets the real-time requirements,which make it suitable for being integration into autonomous driving systems to enhance vehicle perception capacity in complex road environments.

关 键 词：智能交通 小目标检测 YOLOv5 EIoU GAM 

分 类 号：U495[交通运输工程—交通运输规划与管理] TP391.4[交通运输工程—道路与铁道工程]